AI data center power stocks are still beating estimates
Chip stocks fell into a bear market this month. The companies building the physical plant behind AI kept beating their own guidance.
The AI trade split in two this month. Chip stocks fell into a bear market while the companies that build the physical plant behind AI kept beating their own guidance. Vertiv is carrying a $15.0 billion backlog. Eaton reported data center orders up 240 percent. Neither number moved because of anything that happened last week, and that is the point.
Here is what the data says about which half of the AI trade is repricing on sentiment and which half is repricing on contracts.
Why power stocks held while chip stocks fell
Power and electrical names hold up because their revenue is booked years ahead in backlog. Chip stocks reprice on next quarter's demand rumor.
The gap was stark in the week of July 13 to July 20, 2026. The PHLX Semiconductor Index closed about 22 percent below its all time closing high of 14,634.72 set on June 22, 2026. The SOXX ETF is down 18.6 percent in July, its worst month since 2008. The trigger in mid July was Kimi K3, a 2.8 trillion parameter open weights model released free by Moonshot, which made investors ask how much anyone will pay for compute.
The power names did not follow. Intraday on July 20, 2026, Vertiv Holdings (VRT) trades at $299.00, up 1.66 percent. Eaton (ETN) trades at $402.72, up 1.63 percent. Applied Materials (AMAT), which sells chipmaking equipment, trades at $546.00, down 2.66 percent on the same day.
The reason sits on the balance sheet. A transformer ordered today gets delivered in 2028. A GPU order can be cut with a phone call. When the market gets scared about AI demand, it sells whatever can be cancelled fastest.
How big the order books actually are
Order growth is running several times faster than revenue growth at every name in this group. That is the signature of a business that cannot ship as fast as customers want to buy.
Vertiv figure is fiscal Q4 2025 organic orders. Eaton figure is data center orders. GE Vernova is total orders. Source: company filings and earnings calls.
| Company (ticker) | Latest quarter revenue | Backlog | Order growth | Guidance |
|---|---|---|---|---|
| Vertiv (VRT) | $2.65B, up 30.1% | $15.0B, up 109% | Q4 2025 organic orders up 252%, book to bill 2.9x | FY26 sales $13.5B to $14.0B, adjusted EPS $6.30 to $6.40 |
| Eaton (ETN) | $7.45B, up 17% | $22.8B | Data center orders up 240% | FY26 adjusted EPS $13.05 to $13.50 |
| GE Vernova (GEV) | $9.3B, up 16% | Not disclosed | Orders $18.3B, up 71% | FY26 revenue raised to $44.5B to $45.5B |
| Arista (ANET) | Not disclosed | $6.2B deferred revenue | AI revenue target raised to $3.5B | FY26 revenue raised to about $11.5B, up 27.7% |
A book to bill of 2.9 means Vertiv took in nearly three dollars of new orders for every dollar of revenue it recognized.
The Arista line is the one most investors skip past. Deferred revenue is money customers have already committed that the company cannot yet count as revenue. Arista's sits at $6.2 billion, up from $5.37 billion, because supply shortages are delaying delivery. That is demand stuck in a queue, not demand that went away.
Two near term tests: Vertiv reports second quarter results on July 29, 2026 before the open, with consensus adjusted EPS of $1.43. GE Vernova reports Wednesday, July 22, 2026 before the open.
What the compute data says about demand
The Bargo Compute Tightness Index reads 47.5 as of July 20, 2026. That is a balanced regime, and it is tightening, up 2.1 points over 30 days. The index measures how scarce rentable AI compute is by tracking hourly rental prices. Higher means scarcer.
The headline number looks calm. The mix underneath is not.
The H200 sits at 66.6, a tight regime, up 10.9 points in 30 days. The H100 sits at 38.7, loose, and has not moved in 30 days. Current generation capacity is scarce. Prior generation capacity is not.
You can see it in the spot discount, which is how much cheaper a chip gets when a provider rents out spare machines. An H200 renter saves only 25 percent by going spot. An H100 renter saves 56 percent. A narrow spot discount means providers have no idle machines to give away.
Demand is louder still. Token consumption across tracked inference providers hit 58.5 trillion tokens per week as of July 19, 2026, up 32 percent in 30 days and up 149 percent since the series began on May 6. The blended effective price fell to $1.97 per million tokens.
Cheaper per unit, far more units. Every one of those tokens runs on a machine that needs power and cooling. Live series: Compute Tightness Index and Token Demand.
The permitting risk nobody models
On July 14, 2026, New York Governor Kathy Hochul signed an executive order creating the nation's first statewide one year moratorium on new data centers above 50 megawatts, running to July 2027. The stated reason is grid strain and residential electricity costs.
The bull case for this group assumes the buildout happens on schedule. Local politics now decides the schedule.
Morgan Stanley's July 2026 work has hyperscaler compute capacity going from roughly 30 gigawatts in 2025 to about 116.6 gigawatts in 2028, funded by capex of roughly $1.2 trillion in 2027 and $1.4 trillion in 2028. Microsoft is tracking about $190 billion of capex in calendar 2026. Alphabet raised its 2026 capex guidance to $180 billion to $190 billion on April 29, 2026.
Those numbers assume permits, land, interconnection slots and a local government that says yes. New York said no for a year.
Read it two ways. It is a delay risk for anyone selling into New York projects, including Vertiv and the bitcoin miners turned AI hosts such as IREN, at $36.70 intraday on July 20. It is also a scarcity signal. If new capacity gets harder to build, existing capacity gets more valuable, and so does equipment that squeezes more compute out of a site that already has power.
Chips or power, which is the real constraint
Power is the constraint. Chips are available. The capacity to run them is not.
Three pieces of evidence, all as of July 2026, point the same way:
- H100 rentals sit loose at a tightness reading of 38.7, meaning prior generation chips are available and underused
- Arista holds $6.2 billion in deferred revenue because it cannot ship networking gear fast enough
- Eaton and Vertiv report order growth above 240 percent while revenue grows 17 to 30 percent
If chips were the binding constraint, rental prices would spike across every generation. Instead the current generation is tight, the prior generation is loose, and the physical plant lags both.
The bear case you should know
Two Arista insiders sold heavily in mid July 2026. Co-founder Andy Bechtolsheim sold 300,000 shares on July 14 for $54,231,399, an average of $180.77. CEO Jayshree Ullal sold 234,578 shares on July 10 for $43,907,278, an average of $187.18, after selling roughly $45 million more on July 9. The Form 4 filings available do not indicate whether these were made under prearranged plans, so treat the size as a fact and the motive as unknown. The filings are on SEC EDGAR.
The second risk is arithmetic. Order growth of 252 percent cannot repeat. At some point the year ago comparison gets hard, and a slowdown from extraordinary to merely strong tends to get sold like a miss.
What to watch
- The gap between order growth and revenue growth at Vertiv and Eaton. When it closes, the story changes.
- Vertiv's July 29 print and GE Vernova's July 22 print, for whether order momentum held through the June quarter.
- The split inside the Compute Tightness Index, not the headline number. H200 tight and H100 loose is the signal.
- Whether token volume keeps rising while the price per million falls. That combination means the market is expanding.
- Whether other states follow New York on data center permitting.
This is a research note, not investment advice.
Sources
- Vertiv Holdings investor relations, Q1 2026 results and FY26 guidance — https://investors.vertiv.com
- Eaton investor relations, Q1 2026 results and 10-Q backlog — https://www.eaton.com/us/en-us/company/investor-relations.html
- GE Vernova investor relations, Q1 2026 orders and guidance — https://www.gevernova.com/investors
- Arista Networks investor relations, Q1 2026 results and deferred revenue — https://investors.arista.com
- SEC EDGAR, Form 4 insider filings — https://www.sec.gov/edgar/search/
- Office of the Governor of New York, July 14 2026 executive order — https://www.governor.ny.gov/news
- U.S. Energy Information Administration, electricity data — https://www.eia.gov/electricity/
- Bargo Compute Tightness Index — https://www.bargo.ai/compute-tightness
- Bargo Token Demand Index — https://www.bargo.ai/token-demand
- Bargo GPU rental prices — https://www.bargo.ai/gpu-prices
More research at bargo.ai/research.